Temporal heterogeneous graph neural networks for multimodal rumor detection on social media
Early and reliable rumor detection on social media requires joint modeling of multimodal semantics, user–post interactions, propagation structures, and temporal evolution. Existing methods usually enhance multimodal fusion, propagation modeling, or robustness separately, but they often rely on static feature fusion, homogeneous propagation structures, or weak temporal reasoning, which limits their effectiveness in early-stage and cross-event scenarios. To address these limitations, we propose THGNN-MRD, a temporal heterogeneous graph neural network for multimodal rumor detection. The core idea is to formulate rumor detection as dynamic event-level reasoning over a sequence of temporal heterogeneous graph snapshots. Specifically, each event is represented by posts, comments, users, and visual semantic nodes; relation-aware and time-aware bidirectional message passing is performed within each snapshot; and a Mamba state-space encoder is used to model cross-snapshot propagation evolution. To improve discriminability and robustness, THGNN-MRD further integrates gated multi-view fusion, cross-modal contrastive learning, temporal continuity regularization, and domain-adversarial optimization into a unified end-to-end framework. Experiments on four benchmark datasets, including Weibo, Twitter, FakeSV, and FVC, show that THGNN-MRD consistently outperforms representative state-of-the-art methods. It improves the F1-score by 1.2, 1.0, 1.0, and 1.4 percentage points over the strongest competing methods on the four datasets, respectively. Additional early-detection, ablation, sensitivity, efficiency, and gate-behavior analyses further confirm the effectiveness and practical applicability of the proposed framework. These results suggest that modeling rumors as evolving multimodal social events provides a principled and effective solution for trustworthy rumor detection.